China’s Embodied Intelligence Advances Towards Human-Level Applications

China’s Embodied Intelligence Advances to “Human-Level” Applications

On March 19, a report from the Science and Technology Daily highlighted significant advancements in China’s humanoid robots, showcasing their journey from clumsy mechanical movement to seamlessly competing in tennis matches against human players. This leap in technology was marked by impressive demonstrations, including challenges in open environments, such as half-marathons.

During a visit to the Beijing Humanoid Robot Innovation Center (a national-local collaborative center for embodied intelligent robotics), the Science and Technology Daily journalist explored training sites and data bases, engaging with industry experts to understand how China’s embodied intelligence has evolved from “mechanical execution” to “autonomous thinking.”

Equipped with a “mini-brain,” the robots boast world-class athletic abilities.

A recent viral video showcased a humanoid robot engaging in a tennis match with a human, performing agile movements, real-time decision-making, and executing multiple high-quality rallies. Developed by Beijing Galaxy General Robotics, this robot signifies China’s leadership in high-dynamic motion control and real-time autonomous decision-making.

“It does not rely on preset actions; it truly learns on site and makes decisions in real-time,” said Zhao Yuli, Chief Strategy Officer at Beijing Galaxy General Robotics. The team’s proprietary intelligent control algorithm, LATENT, equips the robot with a “movement mini-brain” that allows it to autonomously learn movement logic from scattered human motion data without the need for expensive motion capture systems. Actual tests show that the robot achieves a success rate of over 90% in forehand strikes and can maintain rallies for 20 consecutive matches, locking onto a ball traveling at 50 kilometers per hour within 0.1 seconds, displaying both agility and stability.

On March 17, Wang Xingxing, founder of Yushu Technology, made a bold prediction at the Yabuli Forum: by mid-year, China’s robot sprinting speed is expected to surpass that of humans, rivaling Usain Bolt. During the half-marathon training camp delivery ceremony on March 19, Tang Jian, Chief Technology Officer of the Beijing Humanoid Robot Innovation Center, revealed that last year’s “Embodied Tianshi Ultra” set a global record for humanoid robots in the half-marathon at 2 hours, 40 minutes, and 42 seconds, only reaching amateur levels. This year, the aim is to achieve a completion time of around 1 hour, targeting professional athlete standards.

“Running autonomously in an open environment is as challenging as fully autonomous driving,” Tang stated, noting that China’s embodied intelligence is currently at a parallel development stage with the United States. In two key areas—hardware and motion control—China has already established a lead. From stable long-distance marathon running to executing complex routines like those seen in the Spring Festival Gala, domestic robots have reached world-class standards in both stability and smoothness.

Overcoming challenges in unknown scenarios

Spanning nearly 5,000 square meters, the data training base at the Beijing Humanoid Robot Innovation Center has meticulously recreated over 30 scenarios, including home, supermarket, and office environments. This setup allows more than 120 robots to operate simultaneously, forming the largest and most comprehensive training matrix in the country. The “Tianshi” robot series performs precise tasks such as picking avocados, sorting fruits and vegetables, and even simulating diaper changes for infants, honing their “touch and intuition” in realistic physical environments.

“Real machine data collection is crucial for robots to understand the world,” explained Jiang Weilai, head of the data and training base for embodied intelligent robots. The innovation center has established the first domestic industry standard for data collection in embodied intelligence, achieving a data qualification rate of over 95%. The open-source Robomind dataset has been downloaded over 2 million times, delivering tens of thousands of hours of high-quality data to the market, placing it at the forefront of the industry while progressing towards a target of one million hours of high-quality data.

Tang acknowledged that the biggest challenge in the industry currently lies in the lack of generalization capabilities, meaning robots can only perform specific tasks in fixed environments and struggle to adapt flexibly in unfamiliar settings like humans do. The focus for breakthroughs is on data and algorithms: first, expanding data size and diversity to bridge the gap with language models; second, utilizing synthetic data and real-world data collection to enhance data accuracy; and third, strengthening algorithm development to enable robots to successfully complete any task in unknown environments. “Robots do not need to strictly pursue a humanoid appearance; the core is their ability to solve problems in unknown scenarios,” Tang emphasized.

On March 19, the innovation center delivered 15 units of the “Embodied Tianshi 3.0” and “Embodied Tianshi Ultra” to universities such as Peking University, Beijing Institute of Technology, and Huazhong University of Science and Technology, as well as to partners like NIO and China Shipbuilding Group. They also opened the “Huisikao” platform for full-stack technical support, uniting industry, academia, and research through an open-source ecosystem.